diffusers 0.28.0__py3-none-any.whl → 0.28.1__py3-none-any.whl

diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.28.0"
+__version__ = "0.28.1"
 
 from typing import TYPE_CHECKING
 
@@ -82,11 +82,14 @@ else:
             "ConsistencyDecoderVAE",
             "ControlNetModel",
             "ControlNetXSAdapter",
+            "DiTTransformer2DModel",
+            "HunyuanDiT2DModel",
             "I2VGenXLUNet",
             "Kandinsky3UNet",
             "ModelMixin",
             "MotionAdapter",
             "MultiAdapter",
+            "PixArtTransformer2DModel",
             "PriorTransformer",
             "StableCascadeUNet",
             "T2IAdapter",
@@ -227,6 +230,7 @@ else:
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
             "CycleDiffusionPipeline",
+            "HunyuanDiTPipeline",
             "I2VGenXLPipeline",
             "IFImg2ImgPipeline",
             "IFImg2ImgSuperResolutionPipeline",
@@ -484,11 +488,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ConsistencyDecoderVAE,
             ControlNetModel,
             ControlNetXSAdapter,
+            DiTTransformer2DModel,
+            HunyuanDiT2DModel,
             I2VGenXLUNet,
             Kandinsky3UNet,
             ModelMixin,
             MotionAdapter,
             MultiAdapter,
+            PixArtTransformer2DModel,
             PriorTransformer,
             T2IAdapter,
             T5FilmDecoder,
@@ -607,6 +614,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             CLIPImageProjection,
             CycleDiffusionPipeline,
+            HunyuanDiTPipeline,
             I2VGenXLPipeline,
             IFImg2ImgPipeline,
             IFImg2ImgSuperResolutionPipeline,

diffusers/configuration_utils.py

@@ -706,3 +706,20 @@ def flax_register_to_config(cls):
 
     cls.__init__ = init
     return cls
+
+
+class LegacyConfigMixin(ConfigMixin):
+    r"""
+    A subclass of `ConfigMixin` to resolve class mapping from legacy classes (like `Transformer2DModel`) to more
+    pipeline-specific classes (like `DiTTransformer2DModel`).
+    """
+
+    @classmethod
+    def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs):
+        # To prevent depedency import problem.
+        from .models.model_loading_utils import _fetch_remapped_cls_from_config
+
+        # resolve remapping
+        remapped_class = _fetch_remapped_cls_from_config(config, cls)
+
+        return remapped_class.from_config(config, return_unused_kwargs, **kwargs)

diffusers/models/__init__.py

@@ -36,6 +36,9 @@ if is_torch_available():
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
+    _import_structure["transformers.dit_transformer_2d"] = ["DiTTransformer2DModel"]
+    _import_structure["transformers.hunyuan_transformer_2d"] = ["HunyuanDiT2DModel"]
+    _import_structure["transformers.pixart_transformer_2d"] = ["PixArtTransformer2DModel"]
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
@@ -73,7 +76,10 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .transformers import (
+            DiTTransformer2DModel,
             DualTransformer2DModel,
+            HunyuanDiT2DModel,
+            PixArtTransformer2DModel,
             PriorTransformer,
             T5FilmDecoder,
             Transformer2DModel,

diffusers/models/activations.py

@@ -50,6 +50,18 @@ def get_activation(act_fn: str) -> nn.Module:
         raise ValueError(f"Unsupported activation function: {act_fn}")
 
 
+class FP32SiLU(nn.Module):
+    r"""
+    SiLU activation function with input upcasted to torch.float32.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        return F.silu(inputs.float(), inplace=False).to(inputs.dtype)
+
+
 class GELU(nn.Module):
     r"""
     GELU activation function with tanh approximation support with `approximate="tanh"`.

diffusers/models/attention_processor.py

@@ -103,6 +103,7 @@ class Attention(nn.Module):
         upcast_softmax: bool = False,
         cross_attention_norm: Optional[str] = None,
         cross_attention_norm_num_groups: int = 32,
+        qk_norm: Optional[str] = None,
         added_kv_proj_dim: Optional[int] = None,
         norm_num_groups: Optional[int] = None,
         spatial_norm_dim: Optional[int] = None,
@@ -161,6 +162,15 @@ class Attention(nn.Module):
         else:
             self.spatial_norm = None
 
+        if qk_norm is None:
+            self.norm_q = None
+            self.norm_k = None
+        elif qk_norm == "layer_norm":
+            self.norm_q = nn.LayerNorm(dim_head, eps=eps)
+            self.norm_k = nn.LayerNorm(dim_head, eps=eps)
+        else:
+            raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None or 'layer_norm'")
+
         if cross_attention_norm is None:
             self.norm_cross = None
         elif cross_attention_norm == "layer_norm":
@@ -1426,6 +1436,104 @@ class AttnProcessor2_0:
         return hidden_states
 
 
+class HunyuanAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
+    used in the HunyuanDiT model. It applies a s normalization layer and rotary embedding on query and key vector.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        from .embeddings import apply_rotary_emb
+
+        residual = hidden_states
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            if not attn.is_cross_attention:
+                key = apply_rotary_emb(key, image_rotary_emb)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
 class FusedAttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). It uses

diffusers/models/embeddings.py

@@ -16,10 +16,11 @@ from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 from torch import nn
 
 from ..utils import deprecate
-from .activations import get_activation
+from .activations import FP32SiLU, get_activation
 from .attention_processor import Attention
 
 
@@ -135,6 +136,7 @@ class PatchEmbed(nn.Module):
         flatten=True,
         bias=True,
         interpolation_scale=1,
+        pos_embed_type="sincos",
     ):
         super().__init__()
 
@@ -156,10 +158,18 @@ class PatchEmbed(nn.Module):
         self.height, self.width = height // patch_size, width // patch_size
         self.base_size = height // patch_size
         self.interpolation_scale = interpolation_scale
-        pos_embed = get_2d_sincos_pos_embed(
-            embed_dim, int(num_patches**0.5), base_size=self.base_size, interpolation_scale=self.interpolation_scale
-        )
-        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+        if pos_embed_type is None:
+            self.pos_embed = None
+        elif pos_embed_type == "sincos":
+            pos_embed = get_2d_sincos_pos_embed(
+                embed_dim,
+                int(num_patches**0.5),
+                base_size=self.base_size,
+                interpolation_scale=self.interpolation_scale,
+            )
+            self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+        else:
+            raise ValueError(f"Unsupported pos_embed_type: {pos_embed_type}")
 
     def forward(self, latent):
         height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size
@@ -169,6 +179,8 @@ class PatchEmbed(nn.Module):
             latent = latent.flatten(2).transpose(1, 2)  # BCHW -> BNC
         if self.layer_norm:
             latent = self.norm(latent)
+        if self.pos_embed is None:
+            return latent.to(latent.dtype)
 
         # Interpolate positional embeddings if needed.
         # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160)
@@ -187,6 +199,113 @@ class PatchEmbed(nn.Module):
         return (latent + pos_embed).to(latent.dtype)
 
 
+def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
+    """
+    RoPE for image tokens with 2d structure.
+
+    Args:
+    embed_dim: (`int`):
+        The embedding dimension size
+    crops_coords (`Tuple[int]`)
+        The top-left and bottom-right coordinates of the crop.
+    grid_size (`Tuple[int]`):
+        The grid size of the positional embedding.
+    use_real (`bool`):
+        If True, return real part and imaginary part separately. Otherwise, return complex numbers.
+
+    Returns:
+        `torch.Tensor`: positional embdding with shape `( grid_size * grid_size, embed_dim/2)`.
+    """
+    start, stop = crops_coords
+    grid_h = np.linspace(start[0], stop[0], grid_size[0], endpoint=False, dtype=np.float32)
+    grid_w = np.linspace(start[1], stop[1], grid_size[1], endpoint=False, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)  # [2, W, H]
+
+    grid = grid.reshape([2, 1, *grid.shape[1:]])
+    pos_embed = get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=use_real)
+    return pos_embed
+
+
+def get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=False):
+    assert embed_dim % 4 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_rotary_pos_embed(embed_dim // 2, grid[0].reshape(-1), use_real=use_real)  # (H*W, D/4)
+    emb_w = get_1d_rotary_pos_embed(embed_dim // 2, grid[1].reshape(-1), use_real=use_real)  # (H*W, D/4)
+
+    if use_real:
+        cos = torch.cat([emb_h[0], emb_w[0]], dim=1)  # (H*W, D/2)
+        sin = torch.cat([emb_h[1], emb_w[1]], dim=1)  # (H*W, D/2)
+        return cos, sin
+    else:
+        emb = torch.cat([emb_h, emb_w], dim=1)  # (H*W, D/2)
+        return emb
+
+
+def get_1d_rotary_pos_embed(dim: int, pos: Union[np.ndarray, int], theta: float = 10000.0, use_real=False):
+    """
+    Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
+
+    This function calculates a frequency tensor with complex exponentials using the given dimension 'dim' and the end
+    index 'end'. The 'theta' parameter scales the frequencies. The returned tensor contains complex values in complex64
+    data type.
+
+    Args:
+        dim (`int`): Dimension of the frequency tensor.
+        pos (`np.ndarray` or `int`): Position indices for the frequency tensor. [S] or scalar
+        theta (`float`, *optional*, defaults to 10000.0):
+            Scaling factor for frequency computation. Defaults to 10000.0.
+        use_real (`bool`, *optional*):
+            If True, return real part and imaginary part separately. Otherwise, return complex numbers.
+
+    Returns:
+        `torch.Tensor`: Precomputed frequency tensor with complex exponentials. [S, D/2]
+    """
+    if isinstance(pos, int):
+        pos = np.arange(pos)
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))  # [D/2]
+    t = torch.from_numpy(pos).to(freqs.device)  # type: ignore  # [S]
+    freqs = torch.outer(t, freqs).float()  # type: ignore   # [S, D/2]
+    if use_real:
+        freqs_cos = freqs.cos().repeat_interleave(2, dim=1)  # [S, D]
+        freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
+        return freqs_cos, freqs_sin
+    else:
+        freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64     # [S, D/2]
+        return freqs_cis
+
+
+def apply_rotary_emb(
+    x: torch.Tensor,
+    freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+    to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+    reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+    tensors contain rotary embeddings and are returned as real tensors.
+
+    Args:
+        x (`torch.Tensor`):
+            Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+        freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+    """
+    cos, sin = freqs_cis  # [S, D]
+    cos = cos[None, None]
+    sin = sin[None, None]
+    cos, sin = cos.to(x.device), sin.to(x.device)
+
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+    out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+    return out
+
+
 class TimestepEmbedding(nn.Module):
     def __init__(
         self,
@@ -507,6 +626,88 @@ class CombinedTimestepLabelEmbeddings(nn.Module):
         return conditioning
 
 
+class HunyuanDiTAttentionPool(nn.Module):
+    # Copied from https://github.com/Tencent/HunyuanDiT/blob/cb709308d92e6c7e8d59d0dff41b74d35088db6a/hydit/modules/poolers.py#L6
+
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim + 1, embed_dim) / embed_dim**0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.permute(1, 0, 2)  # NLC -> LNC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (L+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (L+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1],
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+        return x.squeeze(0)
+
+
+class HunyuanCombinedTimestepTextSizeStyleEmbedding(nn.Module):
+    def __init__(self, embedding_dim, pooled_projection_dim=1024, seq_len=256, cross_attention_dim=2048):
+        super().__init__()
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.pooler = HunyuanDiTAttentionPool(
+            seq_len, cross_attention_dim, num_heads=8, output_dim=pooled_projection_dim
+        )
+        # Here we use a default learned embedder layer for future extension.
+        self.style_embedder = nn.Embedding(1, embedding_dim)
+        extra_in_dim = 256 * 6 + embedding_dim + pooled_projection_dim
+        self.extra_embedder = PixArtAlphaTextProjection(
+            in_features=extra_in_dim,
+            hidden_size=embedding_dim * 4,
+            out_features=embedding_dim,
+            act_fn="silu_fp32",
+        )
+
+    def forward(self, timestep, encoder_hidden_states, image_meta_size, style, hidden_dtype=None):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, 256)
+
+        # extra condition1: text
+        pooled_projections = self.pooler(encoder_hidden_states)  # (N, 1024)
+
+        # extra condition2: image meta size embdding
+        image_meta_size = get_timestep_embedding(image_meta_size.view(-1), 256, True, 0)
+        image_meta_size = image_meta_size.to(dtype=hidden_dtype)
+        image_meta_size = image_meta_size.view(-1, 6 * 256)  # (N, 1536)
+
+        # extra condition3: style embedding
+        style_embedding = self.style_embedder(style)  # (N, embedding_dim)
+
+        # Concatenate all extra vectors
+        extra_cond = torch.cat([pooled_projections, image_meta_size, style_embedding], dim=1)
+        conditioning = timesteps_emb + self.extra_embedder(extra_cond)  # [B, D]
+
+        return conditioning
+
+
 class TextTimeEmbedding(nn.Module):
     def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
         super().__init__()
@@ -793,11 +994,18 @@ class PixArtAlphaTextProjection(nn.Module):
     Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
     """
 
-    def __init__(self, in_features, hidden_size, num_tokens=120):
+    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh"):
         super().__init__()
+        if out_features is None:
+            out_features = hidden_size
         self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
-        self.act_1 = nn.GELU(approximate="tanh")
-        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True)
+        if act_fn == "gelu_tanh":
+            self.act_1 = nn.GELU(approximate="tanh")
+        elif act_fn == "silu_fp32":
+            self.act_1 = FP32SiLU()
+        else:
+            raise ValueError(f"Unknown activation function: {act_fn}")
+        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=out_features, bias=True)
 
     def forward(self, caption):
         hidden_states = self.linear_1(caption)

diffusers/models/model_loading_utils.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import importlib
 import inspect
 import os
 from collections import OrderedDict
@@ -32,6 +33,13 @@ from ..utils import (
 
 logger = logging.get_logger(__name__)
 
+_CLASS_REMAPPING_DICT = {
+    "Transformer2DModel": {
+        "ada_norm_zero": "DiTTransformer2DModel",
+        "ada_norm_single": "PixArtTransformer2DModel",
+    }
+}
+
 
 if is_accelerate_available():
     from accelerate import infer_auto_device_map
@@ -61,6 +69,26 @@ def _determine_device_map(model: torch.nn.Module, device_map, max_memory, torch_
     return device_map
 
 
+def _fetch_remapped_cls_from_config(config, old_class):
+    previous_class_name = old_class.__name__
+    remapped_class_name = _CLASS_REMAPPING_DICT.get(previous_class_name).get(config["norm_type"], None)
+
+    # Details:
+    # https://github.com/huggingface/diffusers/pull/7647#discussion_r1621344818
+    if remapped_class_name:
+        # load diffusers library to import compatible and original scheduler
+        diffusers_library = importlib.import_module(__name__.split(".")[0])
+        remapped_class = getattr(diffusers_library, remapped_class_name)
+        logger.info(
+            f"Changing class object to be of `{remapped_class_name}` type from `{previous_class_name}` type."
+            f"This is because `{previous_class_name}` is scheduled to be deprecated in a future version. Note that this"
+            " DOESN'T affect the final results."
+        )
+        return remapped_class
+    else:
+        return old_class
+
+
 def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[str] = None):
     """
     Reads a checkpoint file, returning properly formatted errors if they arise.

diffusers/models/modeling_outputs.py

@@ -15,3 +15,17 @@ class AutoencoderKLOutput(BaseOutput):
     """
 
     latent_dist: "DiagonalGaussianDistribution"  # noqa: F821
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: "torch.Tensor"  # noqa: F821

diffusers/models/modeling_utils.py

@@ -42,7 +42,11 @@ from ..utils import (
     is_torch_version,
     logging,
 )
-from ..utils.hub_utils import PushToHubMixin, load_or_create_model_card, populate_model_card
+from ..utils.hub_utils import (
+    PushToHubMixin,
+    load_or_create_model_card,
+    populate_model_card,
+)
 from .model_loading_utils import (
     _determine_device_map,
     _load_state_dict_into_model,
@@ -1039,3 +1043,55 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             del module.key
             del module.value
             del module.proj_attn
+
+
+class LegacyModelMixin(ModelMixin):
+    r"""
+    A subclass of `ModelMixin` to resolve class mapping from legacy classes (like `Transformer2DModel`) to more
+    pipeline-specific classes (like `DiTTransformer2DModel`).
+    """
+
+    @classmethod
+    @validate_hf_hub_args
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
+        # To prevent depedency import problem.
+        from .model_loading_utils import _fetch_remapped_cls_from_config
+
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", None)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+
+        # Load config if we don't provide a configuration
+        config_path = pretrained_model_name_or_path
+
+        user_agent = {
+            "diffusers": __version__,
+            "file_type": "model",
+            "framework": "pytorch",
+        }
+
+        # load config
+        config, _, _ = cls.load_config(
+            config_path,
+            cache_dir=cache_dir,
+            return_unused_kwargs=True,
+            return_commit_hash=True,
+            force_download=force_download,
+            resume_download=resume_download,
+            proxies=proxies,
+            local_files_only=local_files_only,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            **kwargs,
+        )
+        # resolve remapping
+        remapped_class = _fetch_remapped_cls_from_config(config, cls)
+
+        return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs)

diffusers/models/normalization.py

@@ -176,7 +176,8 @@ class AdaLayerNormContinuous(nn.Module):
             raise ValueError(f"unknown norm_type {norm_type}")
 
     def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
-        emb = self.linear(self.silu(conditioning_embedding))
+        # convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT)
+        emb = self.linear(self.silu(conditioning_embedding).to(x.dtype))
         scale, shift = torch.chunk(emb, 2, dim=1)
         x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
         return x

diffusers/models/transformers/__init__.py

@@ -2,7 +2,10 @@ from ...utils import is_torch_available
 
 
 if is_torch_available():
+    from .dit_transformer_2d import DiTTransformer2DModel
     from .dual_transformer_2d import DualTransformer2DModel
+    from .hunyuan_transformer_2d import HunyuanDiT2DModel
+    from .pixart_transformer_2d import PixArtTransformer2DModel
     from .prior_transformer import PriorTransformer
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel